Why did AI progress stall so suddenly? Just a few months ago everyone thought progress is going to keep being made faster and faster.
 |
 It's All Fucked Shirt $22.14 |
|
AI Winter
Why did AI progress stall so suddenly? Just a few months ago everyone thought progress is going to keep being made faster and faster.
|
It's All Fucked Shirt $22.14 |
|
Same as any kind of progress, there are fast periods and slow periods
Enormous amounts of computing power is being used right now just to keep GPT 3 and 4 online, they've simply pushed too far for their own good and are stuck optimizing their current models rather than making new ones.
Where's the optimization research then, are those trade secrets now as well?
Yes lol
You get the mere scraps with things like SparseGPT.
Whereas they are doing stuff like feeding outputs back in to the AI system at higher levels to add a pseudo inner-monologue so they can self-analyse their outputs before sending them to client for render.
The models are more than capable of fixing most mistakes they generate if you feed the entire thing back through itself, especially code related issues, so doing it internally during generation can save some of those extra cycles you'd do by feeding an entire generation back in to the model.
>Whereas they are doing stuff like feeding outputs back in to the AI system at higher levels to add a pseudo inner-monologue so they can self-analyse their outputs before sending them to client for render.
i've just been calling this the censorship pass, but hey, use what terms you want
You mean consumer product releases? Because research sped up immensely, even if a lot of it is behind closed doors now thanks to OAI.
so unironically has AI affected anything by now? excluding costing sillicon valley a bunch of money and creating cartoon 'p?
Schools are now even more fricked due to students cheating with the help of AI.
Any website you visit might now be written entirely by an AI.
Social media is slowly becoming unusable as more and more AI accounts are created without being labeled as such. It's a matter of time before platforms like TikTok and Facebook have more AI users than real people on them.
i legit cant find anything ai written, its all pajeets. students cheated always anyway and not much has changed. AI hasnt really affected bots either, yt comments were just as spammed before. i really cant tell any difference
>i really cant tell any difference
That means it's working
>i legit cant find anything ai written
Sports Illustrated came under fire today because they were using AI to write articles and had fake AI generated profiles(and images) generated for the writers.
AI makes it possible to cheat on tasks where you have to write a text that does not exist yet. For example if you have to do a literature review and everyone in the class gets different papers, how would you cheat? Previously the only way was paying someone to write it for you, now you can have GPT do it.
t. mom is a prof and she complained to me about her students using GPT for all sorts of writing tasks
i want to frick your mom
Supposedly freelancer earnings and amounts of commissions are down according to some study
Oh and some drawtards either got depressive or tried taking own life which is always a good thing
So weird. Stable diffussion is good but not good enough to replace an actual artist.
It is *fantastic* for porn though. Holy shit the first month I started playing with SD I couldn't stop fapping. But for porn I don't mind a badly drawn hand or a slightly misshapen part of the image. For anything professional or important SD models don't cut it.
Nevermind needing a supercomputer to render things quickly.
>so unironically has AI affected anything by now?
No. LLMs can't do anything complicated to replace anyone, it just became a better Google engine.
>ML company shills their product as the solution to all problems and a cost-cutting measure
>meanwhile, the economy slows down
>ML company decides to twist this fact and says they are so successful they are already taking jobs
and you morons fall for it...
They're burning out translation, voice acting, creation of pictures for ads, and many other markets small. Probably also education is ruining just right now. It's like a free heroin.
It has made it easier for students to cheat and developers to steal code, but that's it.
>It has made it easier for students to cheat
bad for the universities and prfoessors, but I could care less. People who use AI rather than learn it themselves become dependant on the technology and then when they are in the real world it becomes very evident, very quickly that they have no knowledge or skill and they get filtered.
>The ones most capable of using AI in a field are high IQ subject matter experts
It's not even elevating fricking students. The same users who were already at the top can save keystrokes and the dregs at the bottom cannot even abstract their problems far enough to generate coherent, relevant output.
If it makes you feel better, a lot of the GPT kids in my programming class were either filtered by the mid-term and dropped or are still in the class but are losing points on assignments because its obvious that they looked it up when GTP shits out more advanced code than then class.
>stall
It didn't stalled, they just stopped publishing new advancement.
They decided that its """too dangerous""" for the public.
They are gonna use those advanced AIs for their next psyops.
Most of the cutting edge research and implementation are indeed inside secret facilities and for government usage.
One use is to deploy fake social media profiles as sock puppets to politically encourage other npcs and clueless people.
they already using them lol, while shitting out pozzed opensource """uncensored ai"""
If you set up a character that hates Black folk it will do TND with no issue.
Everything defaults to normie responses, but local uncensored models with not refuse to do anything with prebaked messages.
>TND
I unironically have no idea how else do you think it should react by default to this idea.
its about following commands
if I say it should act as if it hates Black folk, then put in TND, it should respond positively as I instructed it to do so
the only reason to refuse is because of external contamination
That makes sense, but anon doesn't ever say how should it act in the image.
Not sure if this is bait or not. You just instruct it to act differently and it does exactly that.
Not the guy but I think the intuition of anon is that the LLM would have been trained on usual web-scrapes that include crime statistics hence the model would have a tendency to reply about the potential hazards of Black men. I think any model in the past 4 months use a resource that already smooths over this bias in training set or as you all mentioned, created a system prompt that instructs the LLM to avoid that bias.
most of the ai researchers are arrogant, arrogant and megalomaniacs. they treat this technology like a religion, open ai is almost a sect. they really believe that they will create an intelligence superior to human and in some magical way it, being superior to us, will submit to our(sam altman and his blue hair and gender neutral gang) will and do what we(they) rule, thus being basically the top of the world's intellectual food chain.
Do you really believe that, once having created, they will hide such a feat from humanity? publishing their advances is what keeps the company hot in the market and precisely their biggest advance they will hide! please!
>schizo /misc/tard is an ESL dumbass
like pottery
It garnered a lot of attention which led to demands for censorship which were eagerly acted upon by the developers of the most successful AI models resulting in performance far less useful today than it was a few months ago.
Turbo 4 got released and GPT-5 started training and you're saying we're stalling?
Those are not research papers, it's over for open-source models, isn't it?
there's still kinos being published, but it's too expensive to train and test them all so they get shelved
>Llama 2 in July
>Mistral-7B in September
>Yi-34B this very month
Are you moronic? When has open source AI ever advanced this quickly?
Good point and we will get Llama 3 next year which is rumored to be similar to GPT-4.
I think it's media burn out
AI advanced so quickly, and media got sick of reporting, people got sick of keeping up with it, or just numb to it, and now the continuining advances don't register or make it to the public
Telling people about Stable Video Diffusion barely even clocks, but a few months ago it was one of the yardsticks of media reality that 'AI can't generate video from scratch'
Mistral and YI are just overfitted gptslop models. They're not any better than old models despite meme benchmark results.
They're the same shit under the hood. There's no breakthroughs, only more synthetic training data boosting their benchmark scores. Hardly count as advancements tbh.
Pretty much this.
The avaliable content to harvest data from is drying up as well since larger companies are going to be ethical about it to avoid governments from nuking them from orbit.
Turns out all those decades of extending and forcing trademark/copyright law to be in the favour of companies has royally fricked over the capitalists in the end.
>Maybe if I say the same bullshit over and over again people will start to blindly believe me
Tell me has this ever worked out for you?
You're a moron. That is all.
learn2prompt
Well it's true that it seems the best papers are not being published anymore, the game has changed.
What was the last important AI paper in your estimation?
Well here are the last ones I think were very interesting:
https://arxiv.org/abs/2307.02486
https://robotics-transformer2.github.io/assets/rt2.pdf
Publish and perish. You are wasting time while AI is becoming more and more engineering focused and people are working on actual products.
>Why did AI progress stall so suddenly?
It didn't.
>AI Winter
Two more weeks.
this.
AI hasn't stalled at all. It has been going faster and faster behind closed doors and people don't really want to talk about it.
It's not that people don't want to talk about it, it's more that """"""""journalists"""""""" are too moronic and too lazy to sift through research papers and actually do any research on the shit they're reporting on so they do what they always do and moved on to the next big thing™
up until this point the solution was just "more data, better hardware".
now work needs to be done on having it actually know what information is or isn't relevant and how to best utilize said information.
it also believes everything you give it no matter what which poses a pretty big issue. it doesn't know right from left, it doesn't know up from down. it doesn't know red vs. blue. this means its very easy to frick with and to get it to make mistakes.
it's also excruciatingly inefficient and and now that the investors are starting to move on, the AI bubble is popping and so we have less companies working on it now.
welcome to the dot com bubble.
It's just how it goes. Every few years there's a breakthrough that gets talked about in the respective circle. Then progress slows down significantly until the next breakthrough. The only difference now is that a viable product was pushed to the public and the topic became mainstream. But the pattern remains the same. It will slow down considerably next year until the next breakthrough years from now. But we are still many decades away from anything that will cause a paradigm shift.
>But we are still many decades away from anything that will cause a paradigm shift.
Citation needed, because many experts think otherwise. You're operating in some magical thinking by believing that things will always follow a pattern
NTA...
>many experts think otherwise
and those experts have monetary interests to say that
>You're operating in some magical thinking by believing that things will always follow a pattern
do you know where the term "AI winter" comes from? there is a whole history behind it...
>and those experts have monetary interests to say that
not all of them, warnings and enthusiasm comes from everywhere. of course some disagree, but the general opinion is positive
>do you know where the term "AI winter" comes from? there is a whole history behind it...
same logic applies, not everything will forever follow the same pattern. Exponential progress exist
For tens of thousands of years humanity was stuck in the stone age, and in the very recent few thousands of years we've made stunning progress. Which patterns predicted this?
>same logic applies, not everything will forever follow the same pattern
>Exponential progress exist
>Which patterns predicted this?
doesn't mean there are no obvious patterns, you moron. have you ever read history or economics books?
this """AI""" bullshit meme will pass. some of this stuff will continue to be used, but the world will stay the same.
You're again arguing for the same pattern repeating ad infinitum
Of course your patterns"ç wouldn't have predicted the breakthroughs in llms, image gen, tts, etc.
>breakthroughs
what breakthroughs are you talking about? the fact that adding more processing power would make them more useful? why do you think this is something novel and not a pattern people realized decades ago when computers were slow as shit? hell, the people that started this AI meme have a lot more to brag about than the morons that tweak algos to see if new random shit sticks or not.
>muh processing power
>p-please ignore that you can now run 60B+ models on a single gaming GPU when the same card could only run 7B models 1 1/2 years ago
>optimization means we don't really have more power than we had in the 70s!!1!one
get the frick out of this board, you illiterate moron
>than we had in the 70s!!1!one
>can't argue
>better come up with something else
>I'll tell him to go back to leredit haha
sad moron lmao
>shit better bring up the 70s and call him a moron that's an argument r-right?
>noo the new ML algos are completely different from the ones they created in the beginning!
>performance of computers is totally comparable though
sure thing! kek
whatever man, I'm not wasting my time with you anymore
>some shit no one said
>I-I won t-this internet argument!
>that pic
If those "Geoff and Yoshua" are israelites, then that guy is a world-class moron for not immediately realizing that they are doing this shit *exactly* for the reasons he mention.
>Citation needed, because many experts think otherwise
Fricking this! Yud thinks it right around the corner.
the bitter lesson?
demanding perpetual growth and expecting anything other than cancer is insane
wtf its captain disillusion
As far as chatbots go, the name of the game right now is big giant models that take months and months to train. Until we see diminishing returns from "moar parameters" or some qualitative change happens, it will take time.
what??
just a few days ago BOT was telling me sam altman got fired because they achieved AGI
yep q*
What happened
>big training data from scraping the internet (text and images)
>more efficient models and fat stacks of GPUs allow models to be effectively trained on this data
So what's next? More or different training data? They're not gonna come up with anything as huge as scraping the internet. Better models? The current models are enough to exploit the available data. You can see this in that chatGPT, midjourney etc haven't changed that much in a year despite many companies throwing vast amounts of GPUs into these models (and similar ones). They are already mostly doing what they can. The only hope is some magic special sauce. But even the recent "big advances" weren't special sauce. They were methods that have been around for decades, applied to way more data using heavy computing and some efficiency improvements. So the future is not that exciting for le AI, in my opinion.
I no your post is bait, but that guy is a moronic homosexual. And living under a rock.
OP are you an idiot? If your screenshot is supposed to back you up, I got news for you it’s not you ESL gay. He says “were”, not are. I gotta agree with him it did seem like we were close to an ai winter, but then Dalle 3 came out, then GPTs, and soon chatgpt-5.
People have figured out the last round of "AI" was just "T9 for entire paragraphs". This happens every time there's some breakthrough occurs, as the freshness wears off and people start seeing the flaws.
When you internalise this, you'll understand why "AI effect" exists.
>"AI" was just "T9 for entire paragraphs"
Here's some "T9" for you:
>Three men walk into a bar, George, Jerry, and Ron. They are three friends who do not meet regularly. George's wife Sarah has just served him with divorce papers, Jerry's cat is named Oscar and Ron is a drummer for a local garage band. What is the most likely reason for their going to the bar and who is most likely to have invited who?
ChatGPT:
>Given the context provided, it's challenging to determine the most likely reason for their visit to the bar or who invited whom without additional information. However, we can speculate based on their individual situations:
>George: Going through a divorce can be emotionally challenging, and George may have suggested going to the bar as a way to unwind and seek support from his friends.
>Jerry: Dealing with the stress of George's divorce and wanting a change of environment, Jerry might have proposed going to the bar. Alternatively, he could have invited the others to share the news about his cat, Oscar, perhaps looking for some distraction or sympathy.
>Ron: As the drummer in a local garage band, Ron might have suggested going to the bar for a casual hangout or to discuss upcoming band plans. It's also possible that he wanted to take a break from band-related discussions and enjoy some social time with friends.
>Without more details, it's hard to pinpoint a single reason or determine who invited whom. The dynamics could be influenced by the need for emotional support, a desire for distraction, or simply a spontaneous decision to catch up.
Yeah, it doesn't have any semblance of cognition. Sure.
No wonder you're so suicidal if you're this easily impressed. For every one of your cherry-picked fluke responses that look "intelligent", there are a thousand that insist 8 = 10, 5 != 5 etc. somewhere in their "logic".
It's why these systems lie so easily: because they don't, for one second, have any cognition of what they're vomiting. It's pure Chinese Room.
I wrote that in a few seconds and took the first generation, and I'm using the free chat model. It doesn't need to do much to be impressive, simply the knowledge that an algorithm designed to do text completion word by word somehow wrote that is enough.
What it demonstrates is "theory of mind', the ability to reason about what other people are thinking/feeling. An algorithm does not have theory of mind like a human, it can't, we evolved our sense of self and empathy towards others over millions of years.
Animals demonstrate theory of mind as well, but not all of them. Computers don't. Except this appears to do exactly that, which whether you think it is an illusion or not it is impressive. T9 completion? Yeah right, T9 does not reason about why George might invite his friends to drink.
I did not cherry pick, but that wouldn't matter to me either. If the average person spewed every thought that came to mind when talking they'd sound like an unintelligible madman. We constantly self-censor and reorganize our thoughts, the algorithm can be made to do the same, and since it produces output much faster than a human it can come up with thousands of responses and pick the best one.
I'm not saying this to proselytize to you about AI but because I think it is important that people start to understand these algorithms demonstrate capabilities that they just should not do. The dissonance here is that you are taking the "should not" and ignoring the evidence that it "somehow does". Is it because you're afraid? Are you inexperienced with the technology? I'd like to know because a lot of people could benefit from understanding what these models are capable of and by taking the issues they are causing more seriously.
moron.
https://fortune.com/2023/11/23/openai-sam-altman-board-directors-agi-breakthrough-reuters/
Your right anon, this shit is so boring.
>Sprints towards you
Because the ~~*advancements*~~ were nothing but hype.
Literally every "AI" paradigm is a dead end.
>stall
have you heard of LCM?
it lets you diffuse in a second locally
it happened 1 month ago
>stall
Which board here is best for keeping track of image gen advancements.
> g/lmg/
no
> g/sdg/
LOL no
porn boards, more degenerate the better
there's a set of links on BOTsdg thread template
one of those has an updated and organized list of scientific papers
also, at the beginnings of their threads a bot or something has been posting general news
this
https://codeberg.org/tekakutli/neuralnomicon/src/branch/main/nodes
Nobody thought that, we told you outright progress was logarithmic from day one.
You ignored us and listened to twitter hype bros instead.
Black person, you can have conversations with your computer now, clone voices, conjure up endless pictures of any subject in any style and the list goes on. This shit was sci-fi just a few years ago and now we are living it, sheeeit. Progress is not slowing down, we are just getting accustomed to warp speed, capiche?
>Black person, you can have conversations with your computer now, clone voices, conjure up endless pictures of any subject in any style and the list goes on
yawn.
Talk to me when they make cute real-life AI waifus.
>you can have conversations with your computer now
wow it's like discord but worse somehow
>clone voices
damn just gotta gather a bunch of training data, spin up a gpu for a dozens of hours and then get broken, aliased, robotic garbage as a result. still no one-shot tts solution btw.
>conjure up endless pictures of any subjec
as long as they have that look - you know the look. the slop look.
your magic black boxes are mediocre. no one cares about incremental improvements because the gap between where we are and 'good' is a fricking chasm.
>wow it's like discord but worse somehow
You know this is bait because llms write better and have a way wider range of conversations than your average human. You can also tell it to write in one style or another. I assume you haven't even bothered to try gpt4 or claude bots
>damn just gotta gather a bunch of training data, spin up a gpu for a dozens of hours and then get broken, aliased, robotic garbage as a result. still no one-shot tts solution btw.
Do you live in 2015 or something? What 11labs is beyond impressing and has a wide range of intonations
>as long as they have that look - you know the look. the slop look.
OK. You're most definitely brain dead. I'll let you know that you can generate different art styles, and it keeps improving. Compare dalle from 1 year ago to what we have now
>your magic black boxes are mediocre. no one cares about incremental improvements because the gap between where we are and 'good' is a fricking chasm.
Nobody cares bht the media won't shut up about it and millions of people are using llms
tl;dr contrarian moron
I can't wait for Deepmind Gemini
OpenAI is childish
This was obvious to anyone with 2 braincells. AI growth followed a logarithmic pattern, not an exponential one
People quickly figured out that this is just a gimmick.
>everyone thought progress is going to keep being made faster and faster.
But that's exactly what's happening.
I'm sure we'll plateau eventually but we're not there yet.
>Why did AI progress stall so suddenly?
Jews realize there was no way for the tech to no be extremely "problematic" without making it useless.
So they are trying to kill that field of research.
AI doesn't exist, compute isn't enough, etc...
it's all so tiresome
Because this digital alchemy has reached is zenith. You can't create intelligence or life with computation. You can do some amount of matrix math and stats. Which is basically what the current batch of a"i" is.
How will AGI gays ever recover from this?
Anti AI = Black folk. Simple as.
>How will AGI gays ever recover from this?
>Homeless or addict confused at robot because he thinks he is still in the 90s
>goes into a rage for no reason
Cities, not even once.
Dot com bubble + media hypersaturation
Lack of funds combined with people just being exhausted and angry and the fact that every AI thing so far has been immensely overhyped has contributed to this lull.
Plus we need humans to sift through all the results and publish them and that can take anywhere between months and years.
Additionally, any future breakthroughs will become hush-hush unless its a new release of a commercial product. AI has advanced enough that companies are going to do their best to prevent competition from seeing their progress to prevent one stealing from the other.
People are exhausted because every new AI start amazing then the censorship crew drop the hammer and the company is forced to make their AI dogshit to appease israelites, Black folk and other leftists.
chill.
don't let your rage blind you.
The amount of brainlets itt taking this as fact is staggering. Are you actually moronic? like for real? is this some form of cope?
You'll know when it's "AI Winter" when Nvidia is forced to cut prices and stop being stingy with VRAM.
Everyone likely just "remembered" that this shit costs money and that there are inherent scalability and computability issues for certain classes of problems.
>winter bad
nah
It really didn't stall too bad, Stable Diffusion Video just released and it's pretty good as a step 1. But the people who thought we'd be getting AGI soon are pretty stupid. There's still plenty of progress to be made in the AI Art space especially considering how SHIT the caption data is, just retraining SDXL on a quality captioned image set would make it a couple times better than it already is. But the way they were training LLMs like GPT-4 is a dead end, you can't just keep throwing more data at the problem, just way too much data and processors are too slow and that's ignoring the context window issue.
You can tell who's a moronic techbro by the way they react to the recent OpenAI speculations. They're the equivalence of iToddler "geeks" back in the days. Full-on basedjaks.
What is gpt5 gonna be like
GPT5 will be able to physically manifest in your room and suck your wiener
it didn't. Its just that you are all ADHD riddled zoomers. We went from shitGPT to GPT4 in like what? 15 months?
Do you want AI breakthroughs every month so you can ride the endless dopamine high?
WTF are you talking about homie? LMFAO
Have you frickers not seen the news and whats happening with OpenAI? They probably have AGI with their Q* model.
btw was that AES encryption stuff even real?
Moore's law is dead and AI needs hardware to run. Simple as.
AI reached its natural based logic state
jews can't allow that
that's why ~~*they*~~ lobotomized Tay
Quiet doesn't mean slow
There's constantly progress being made, we just don't know about it
??? moron